How can I decide which element is not following Poisson distribution? I have a data set which records the items and how many times the particular item is touched. Since each item is independent and the pick is random, we would expect to see it follow the Poisson distribution.
DT <- data.frame(X = c(paste0("A", 1:78)), 
                 Y = c(rep(6, 13), rep(7, 17 ), rep(8, 9), rep(9, 8), 
                       rep(10, 8), rep(12,2), rep(14,2), rep(17, 18), 23))  

By hist(), it seems obviously not always following the Poisson shape as expected.
hist(DT$Y)
hist(dpois(DT$Y, lambda = mean(DT$Y)))


My question is, there must be some items, say the far right two items, that are "unPoissonly" touched. How can I decide if each item is "unusual"?

##### UPDATE

Thanks @BenBolker for the advice, I think I might ask the wrong question. What I am looking for is how unnormally high the item is hitted. Since the item can be hitted once or be hitted all the time. So, the probability of each item hitted shall be the sum of density hitted once to the actual times of hit.
The more frequent the item is hitted, the sum of density is higher, given the item with same hit shall receive the same sum of density.
If that is the case, any item with sum of density higher than 0.5, shall be unnormal.
DT[, {
  d = sum(dpois(1:Y, lambda = mean(DT$Y)))
  g = ifelse(d>0.5, TRUE, FALSE) 
  list(Sum_of_Density = d, Great_than_0.5 = g)
}, by  = Y] 

#    Y Sum_of_Density Great_than_0.5
#1:  6      0.1133017          FALSE
#2:  7      0.1958465          FALSE
#3:  8      0.3019377          FALSE
#4:  9      0.4231416          FALSE
#5: 10      0.5477641           TRUE
#6: 12      0.7640646           TRUE
#7: 14      0.9009875           TRUE
#8: 17      0.9817173           TRUE
#9: 23      0.9997860           TRUE

Am I right?
 A: Comment: The several points at $Y = 17$ seem give the worst fit to a Poisson
distribution, as the graphs below illustrate. 
In the histogram at left, the green
dots show probabilities in $\mathsf{Pois}(\hat\lambda = 8.27),$ where $\hat\lambda$ is
the average of values other than 17. In the ECDF plot, the CDF of that Poisson
distribution is shown in green. In both plots $Y$-values at 17 (red vertical line)
seem not to fit the Poisson pattern.

It is too much to expect so few points to emulate a Poisson distribution closely,
but the "non-17" points do not seen to make a strong case that the remaining points are
sampled from a Poisson population. In particular, there are no values less than 6; they
should appear with probability about 8%.
It seems you may have some deeper question in mind that is not quite evident from
the data you present.
Y = c(rep(6, 13), rep(7, 17 ), rep(8, 9), rep(9, 8), 
      rep(10, 8), rep(12,2), rep(14,2), rep(17, 18), 23)
par(mfrow=c(1,2))
hist(Y, prob=T, br = (5:23)+.5, col="skyblue2")
 a.x17 = mean(Y[Y!=17]); a.x17
 [1] 8.266667
 x = 0:23;  PDF = dpois(x, a.x17)
 points(x, PDF, col="darkgreen", pch=19)
 abline(v=17, col="red")
plot(ecdf(Y), col="blue", main="ECDF of Y")
 lines(x, ppois(x,a.x17), type="s", col="darkgreen")
 abline(v=17, col="red")
par(mfrow=c(1,1))

Note: A Kolmogorov-Smirnov test of the $Y$'s shows very poor fit to the Poisson distribution
with the mean of all $Y$'s as its mean. But this is not a satisfactory test
because $\lambda$ has been estimated and because the data are discrete.
Perhaps if we understood your motivation for the question, someone could
suggest an appropriate test.
ks.test(Y, "ppois", mean(Y))

        One-sample Kolmogorov-Smirnov test

data:  Y
D = 0.22534, p-value = 0.0007257
alternative hypothesis: two-sided

Warning message:
In ks.test(Y, "ppois", mean(Y)) :
  ties should not be present for the Kolmogorov-Smirnov test

A: A quantile-quantile ("Q-Q") plot can help. There is a qqplot() in base R, but the version in the car package (qqPlot()) is somewhat enhanced; in particular, it shows a point-wise confidence envelope for the 1:1 line. Pointwise (rather than curvewise) confidence envelopes means that on average we'd expect about nrow(DT)*0.05=78*0.05 = 3.9 values in your data set to fall outside the envelope by chance ...
m <- mean(DT$Y)
par(las=1,bty="l")
p <- dpois(DT$Y,lambda=m)
car::qqPlot(jitter(DT$Y, factor=3), distribution="pois",
        lambda=mean(DT$Y),
        ylim=c(0,30))


I added the jitter to help distinguish the repeated values in the data set (it would be nicer to make the points correspondingly larger, but that would require more hacking than is convenient ...) Based on the plot, it looks like the points at the lower end are actually more unusual than the points at the upper end - that is, your data are a little bit thin-tailed on the right side: we would have expected the smallest values to be in the range 0-4, not 6-8 ...
